Most people first meet AI as a head in a box: a chat window that can answer questions, explain concepts, and produce text on command. That's already useful. But it still leaves the human doing the carrying. You ask, it answers, you do the work.
The real transition happens when the system gets hands. An agent can read a folder, inspect files, run commands, write code, and fix things in place. That's the move to octopus in a box. Not intelligence as conversation. Intelligence as reach.
I don't know whether the world that's coming will feel more liberating or more brutal. I do know that I need to learn how to build in it. I learn by building, by mentoring, and by understanding a little better every day.
This site is the curriculum for that crossing. Not a pile of articles. A path.
The underlying discipline is narrower than the path makes it look. This is really a curriculum in accountability and context: verify what the system did, and give it enough of your world that it can do something specific.
The path
First, replace hype and vague fascination with a working picture of what changed. The goal here is not mastery. It's orientation.
- What's All the Fuss About? — the plain-English onramp
- The Cheating Sheet — the thesis: documents with hands
- Learn by Building — why making clarifies faster than studying
Next, get the tool out of the browser and onto your machine. This is where the system stops being a clever answer box and starts being able to act on files.
- Zero to Developer (Mac), Windows, or Linux — the fastest way to get an agent with hands
- Set Up Your Workspace — the folder becomes the interface
- GitHub and SSH Keys — the minimum infrastructure for real work
Then pick one project shape and ship something. The point is not to become an AI expert. The point is to make one real thing with the new interface.
- What Do You Want to Build? — choose the first shape
- Board Game — build from rules you already know
- Chatbot — turn messy knowledge into a system that answers
Before public means before public. If the thing is going to touch the internet or other people, add a short safety layer: Security for Directors, Prompt Injection, and Before You Deploy.
The goal is not to freeze. The goal is to stop shipping blind.
Once one build exists, the first improvement loop does not require more agents. A single agent can review the artifact it just made, reread the transcript, surface friction, and tighten the result. Another agent helps, but the core move is review plus correction. The artifact teaches the process it needed. The flywheel follows the action.
- Flywheel — turn the residue of work into visible interventions
- The Flywheel Follows the Action — improvement starts from what the work already produced
- Memory Is Files — keep the evidence trail the next pass can read
After the loop works on one project, widen both labor and context. Add more agents, or give the same agent more of your world: more folders, more exports, more history, more evidence. Cross-agent conversation is a cheap scaling primitive. The wall of data is what gives that conversation something larger to work over.
- Wall of Data — gather your own corpus into one place
- Guide-Based Development — make the method portable across agents
- The Context Gold Mine — why wider context changes what the system can do
The loop after the crossing: make one thing, review it, then widen the labor and context around it.
Flywheel is the improvement loop. Guide-Based Development helps that loop survive more agents. Wall of Data gives the loop a bigger world to operate over.
What this curriculum is really teaching
On the surface, it looks like setup instructions, project guides, and a book. Underneath, it's teaching a narrower thing: how to move from reading about intelligence to working with intelligence that can make things, improve those things, and scale them across a bigger body of context.
Read gives you the model. Touch gives the model hands. Make the thing you want to make. Then let the artifact teach you the process it needed. Improve closes the loop through review and correction. Scale widens the labor and the context.
The two constraints underneath all five stages stay the same: accountability keeps the work honest, and context keeps it from collapsing back into generic autocomplete.
If you only stay in the read stage, AI remains news. If you cross into make, improve, and scale, it becomes part of your practice. The process should condense around the artifact, not the other way around.
How to use the site
- If you're curious but skeptical, start with What's All the Fuss About?.
- If the thesis is still fuzzy, read The Cheating Sheet.
- If you're ready to get hands, do Zero to Developer on your platform.
- If you want the first real build, go to What Do You Want to Build? and pick one shape.
- If you are about to expose a build to the internet, do Security for Directors and Before You Deploy first.
- If you want a build to get better, start with Flywheel and review the work that already happened.
- If you want to scale beyond one project, build a Wall of Data and make the method portable.
This is the heart of the mentoring. The site just makes the path repeatable.